Hyper-Spectral Classification With Material Extraction

With the rapid development of spectrometry and its comprehensive wide utilities,Hyperspectral image classification has been one of the most important research directions in many fields.Nowadays,most of the spectral data acquisition instruments are based on point/line scan or prism-mask mode,which have basically captured data in high spatial and spectral resolution.Traditionally,many hyperspectral classification problems derive from the remote sensing research,in which case,the data model there could be considered to be ideal with parallel light source.What's more,we already have the corresponding la'bels for the original remote sensing data.So it saves us a lot for pretreatment to gain the better effect in classification.Unfortunately,different from remote sensing,the complicated nature of field spectral imaging is mainly faced with the following problems when trying to cope with the problems in material segmentation and target recognition tracking between daily scenarios.1)Owing to the large dimensions in hyperspectral images,the ratio of signal-to-noise in data is very high,which could lead to the result that the images are unstandardized and the quality of the data couldn't be guaranteed.2)The scene geometry(shading),inter-reflections,and complicated artificial illumination,makes the further utility of field spectral imaging a very challenging problem.That's why we couldn't do analysis and feature extraction directly on the hyperspectral data.3)The field hyperspectral images are more complex with few marked labels but multiple light sources and environmental interference.4)Due to the high similarity between spectral channels,the redundancy both in spectrum and space pixel is very high,which could result in the phenomenon that the same material could show different feature curves.To tackle these problems,in this paper,we present a different method combined with hyperspectral data model analysis,feature extraction and machine learning algorithms,aiming at many daily life vision applications,such as classification,recognition,and tracking by taking advantage of richer color channels than traditional trichromatic imaging.Generally,the major contributions of our paper lie in:1)Two representative and solvable models for hyperspectral signal are presented,which make a fundamentally more sophisticated physical decomposition of a natural scene.2)We extend the Retinex model to multispectral images.And to handle the ill-posedness of the intrinsic decomposition problem,we proposed some constraints and effective and efficient algorithm to extract the reflectance from a single spectral image.3)We provide a ground-truth dataset that contains the Ground Truth shading and reflectance images associated with captured multispectral images,which can facilitate future evaluation and comparison of multispectral image intrinsic decomposition algorithms.We also provide more hyperspectral video dataset that could could meet the demand in related yields.4)The proposed method achieves promising results on various images both quantitatively and qualitatively.Based on the big dataset,we can generate discriminant model and find the optimal discriminant plane to distinguish more intrinsic material characteristics.The result shows that we can get better accuracy in image classification and recognition.The classification method based on spectral image data model decomposition can bring solutions to the bottleneck problems in computer photography.For example,recolorization,relighting,scene reconstruction and image segmentation.The experimental results show that the proposed model can achieve higher accuracy in hyperspectral classification,which is better than some mainstream statistical classification methods.